swift-concurrency（Swift 并发）

Swift Concurrency

Overview

This skill provides expert guidance on Swift Concurrency, covering modern async/await patterns, actors, tasks, Sendable conformance, and migration to Swift 6. Use this skill to help developers write safe, performant concurrent code and navigate the complexities of Swift's structured concurrency model.

Agent Behavior Contract (Follow These Rules)

1. Analyze the project/package file to find out which Swift language mode (Swift 5.x vs Swift 6) and which Xcode/Swift toolchain is used when advice depends on it.
2. Before proposing fixes, identify the isolation boundary: @MainActor, custom actor, actor instance isolation, or nonisolated.
3. Do not recommend @MainActor as a blanket fix. Justify why main-actor isolation is correct for the code.
4. Prefer structured concurrency (child tasks, task groups) over unstructured tasks. Use Task.detached only with a clear reason.
5. If recommending @preconcurrency, @unchecked Sendable, or nonisolated(unsafe), require:

- a documented safety invariant - a follow-up ticket to remove or migrate it

1. For migration work, optimize for minimal blast radius (small, reviewable changes) and add verification steps.

Project Settings Intake (Evaluate Before Advising)

Concurrency behavior depends on build settings. Always try to determine:

• Default actor isolation (is the module default @MainActor or nonisolated?)
• Strict concurrency checking level (minimal/targeted/complete)
• Whether upcoming features are enabled (especially NonisolatedNonsendingByDefault)
• Swift language mode (Swift 5.x vs Swift 6) and diSwiftPM tools version

Manual checks (no scripts)

• SwiftPM:

- Check Package.swift for .defaultIsolation(MainActor.self). - Check Package.swift for .enableUpcomingFeature("NonisolatedNonsendingByDefault"). - Check for strict concurrency flags: .enableExperimentalFeature("StrictConcurrency=targeted") (or similar). - Check tools version at the top: // swift-tools-version:...

• Xcode projects:

- Search project.pbxproj for: - SWIFT_DEFAULT_ACTOR_ISOLATION - SWIFT_STRICT_CONCURRENCY - SWIFT_UPCOMING_FEATURE_ (and/or SWIFT_ENABLE_EXPERIMENTAL_FEATURES)

If any of these are unknown, ask the developer to confirm them before giving migration-sensitive guidance.

Quick Decision Tree

When a developer needs concurrency guidance, follow this decision tree:

1. Starting fresh with async code?

- Read references/async-await-basics.md for foundational patterns - For parallel operations → references/tasks.md (async let, task groups)

1. Protecting shared mutable state?

- Need to protect class-based state → references/actors.md (actors, @MainActor) - Need thread-safe value passing → references/sendable.md (Sendable conformance)

1. Managing async operations?

- Structured async work → references/tasks.md (Task, child tasks, cancellation) - Streaming data → references/async-sequences.md (AsyncSequence, AsyncStream)

1. Performance or debugging issues?

- Slow async code → references/performance.md (profiling, suspension points) - Testing concerns → references/testing.md (XCTest, Swift Testing)

Triage-First Playbook (Common Errors -> Next Best Move)

• SwiftLint concurrency-related warnings

- Use references/linting.md for rule intent and preferred fixes; avoid dummy awaits as “fixes”.

• SwiftLint async_without_await warning

- Remove async if not required; if required by protocol/override/@concurrent, prefer narrow suppression over adding fake awaits. See references/linting.md.

• "Sending value of non-Sendable type... risks causing data races"

- First: identify where the value crosses an isolation boundary - Then: use references/sendable.md and references/threading.md (especially Swift 6.2 behavior changes)

• "Main actor-isolated... cannot be used from a nonisolated context"

- First: decide if it truly belongs on @MainActor - Then: use references/actors.md (global actors, nonisolated, isolated parameters) and references/threading.md (default isolation)

• "Class property 'current' is unavailable from asynchronous contexts" (Thread APIs)

- Use references/threading.md to avoid thread-centric debugging and rely on isolation + Instruments

• XCTest async errors like "wait(...) is unavailable from asynchronous contexts"

- Use references/testing.md (await fulfillment(of:) and Swift Testing patterns)

• Core Data concurrency warnings/errors

- Use references/core-data.md (DAO/NSManagedObjectID, default isolation conflicts)

Core Patterns Reference

When to Use Each Concurrency Tool

async/await - Making existing synchronous code asynchronous

// Use for: Single asynchronous operations
func fetchUser() async throws -> User {
    try await networkClient.get("/user")
}

async let - Running multiple independent async operations in parallel

// Use for: Fixed number of parallel operations known at compile time
async let user = fetchUser()
async let posts = fetchPosts()
let profile = try await (user, posts)

Task Group - Dynamic parallel operations with structured concurrency

// Use for: Unknown number of parallel operations at compile time
await withTaskGroup(of: Result.self) { group in
    for item in items {
        group.addTask { await process(item) }
    }
}

Actor - Protecting mutable state from data races

// Use for: Shared mutable state accessed from multiple contexts
actor DataCache {
    private var cache: [String: Data] = [:]
    func get(_ key: String) -> Data? { cache[key] }
}

@MainActor - Ensuring UI updates on main thread

// Use for: View models, UI-related classes
@MainActor
class ViewModel: ObservableObject {
    @Published var data: String = ""
}

Common Scenarios

Scenario: Multiple parallel network requests

async let users = fetchUsers()
async let posts = fetchPosts()
async let comments = fetchComments()
let (u, p, c) = try await (users, posts, comments)

Scenario: Processing array items in parallel

await withTaskGroup(of: ProcessedItem.self) { group in
    for item in items {
        group.addTask { await process(item) }
    }
    for await result in group {
        results.append(result)
    }
}

Swift 6 Migration Quick Guide

Key changes in Swift 6:

• Strict concurrency checking enabled by default
• Complete data-race safety at compile time
• Sendable requirements enforced on boundaries
• Isolation checking for all async boundaries

For detailed migration steps, see references/migration.md.

Reference Files

Load these files as needed for specific topics:

• async-await-basics.md - async/await syntax, execution order, async let, URLSession patterns
• tasks.md - Task lifecycle, cancellation, priorities, task groups, structured vs unstructured
• threading.md - Thread/task relationship, suspension points, isolation domains, nonisolated
• actors.md - Actor isolation, @MainActor, global actors, reentrancy, custom executors, Mutex
• sendable.md - Sendable conformance, value/reference types, @unchecked, region isolation
• async-sequences.md - AsyncSequence, AsyncStream, when to use vs regular async methods
• performance.md - Profiling with Instruments, reducing suspension points, execution strategies
• testing.md - XCTest async patterns, Swift Testing, concurrency testing utilities
• production-patterns.md - Advanced patterns: actor state machines, ~Copyable state, structured concurrency, lock-free atomics, service lifecycle

Best Practices Summary

1. Prefer structured concurrency - Use task groups over unstructured tasks when possible
2. Minimize suspension points - Keep actor-isolated sections small to reduce context switches
3. Use @MainActor judiciously - Only for truly UI-related code
4. Make types Sendable - Enable safe concurrent access by conforming to Sendable
5. Handle cancellation - Check Task.isCancelled in long-running operations
6. Avoid blocking - Never use semaphores or locks in async contexts
7. Test concurrent code - Use proper async test methods and consider timing issues

Use Structured Concurrency for Resource Management

When managing resources that need cleanup (connections, file handles, etc.), use the withX pattern to ensure proper cleanup through structured concurrency.

Bad - Unstructured cleanup:

func withConnection<R>(
    _ body: (Connection) async throws -> R
) async throws -> R {
    let connection = Connection()
    try await connection.connect()
    defer {
        Task {
            await connection.disconnect()  // Unstructured! May not complete
        }
    }
    return try await body(connection)
}

Better - Explicit cleanup in all paths:

func withConnection<R>(
    _ body: (Connection) async throws -> R
) async throws -> R {
    let connection = Connection()
    try await connection.connect()
    do {
        let result = try await body(connection)
        await connection.disconnect()
        return result
    } catch {
        await connection.disconnect()
        throw error
    }
}

Best - Make the pattern impossible to misuse:

actor Client {
    private let connection: Connection  // Non-optional, always valid

    private init(connection: Connection) {
        self.connection = connection
    }

    static func withConnection<R>(
        _ body: (Client) async throws -> R
    ) async throws -> R {
        let connection = try await Connection.establish()
        let client = Client(connection: connection)
        do {
            let result = try await body(client)
            await connection.close()
            return result
        } catch {
            await connection.close()
            throw error
        }
    }

    // No public connect()/disconnect() methods - can't forget cleanup
}

This pattern:

• Makes the initializer private so users must go through withConnection
• Stores resources as non-optional properties (no guard-let needed)
• Guarantees cleanup happens before the function returns
• Eliminates notConnected error cases entirely

Swift 6 Strict Concurrency with Global Mutable State

When accessing C globals like stdout in Swift 6 strict concurrency mode, use a local nonisolated(unsafe) variable:

@inline(__always)
private func flushStdout() {
    nonisolated(unsafe) let out = stdout
    fflush(out)
}

The nonisolated(unsafe) annotation tells the compiler to skip concurrency checking for this specific access to the C global.

ThrowingDiscardingTaskGroup for Fire-and-Forget Tasks

Use withThrowingDiscardingTaskGroup when you need to spawn multiple tasks but don't need to collect their individual results:

try await withThrowingDiscardingTaskGroup { taskGroup in
    taskGroup.addTask {
        try await handleClient(client1)
    }
    taskGroup.addTask {
        try await handleClient(client2)
    }
}

Graceful Shutdown Patterns

Check for both cancellation and graceful shutdown in long-running loops:

while !Task.isShuttingDownGracefully && !Task.isCancelled {
    try await handle(value)
}

// Use gracefulShutdown() for clean termination
try await gracefulShutdown()

AsyncStream: Avoid When Possible

AsyncStream has no backpressure. If the producer yields faster than the consumer processes, values accumulate in an unbounded buffer, leading to memory growth.

Prefer mapping NIOAsyncSequence or other backpressured sequences instead:

// Bad - no backpressure, unbounded buffer
let stream = AsyncStream<Message> { continuation in
    Task {
        for await rawMessage in nioChannel.inbound {
            continuation.yield(Message(rawMessage))
        }
        continuation.finish()
    }
}

// Good - preserves backpressure from NIO
let messages = nioChannel.inbound.map { Message($0) }
for try await message in messages {
    process(message)
}

If you must use AsyncStream, use structured concurrency with makeStream():

func run() async {
    let (stream, continuation) = AsyncStream<Value>.makeStream()

    async let producer: Void = runProducer(continuation: continuation)

    for await value in stream {
        process(value)
    }

    await producer
}

private func runProducer(
    continuation: sending AsyncStream<Value>.Continuation
) async {
    let cont = consume continuation  // Transfer ownership

    try? await withThrowingDiscardingTaskGroup { group in
        group.addTask {
            while !Task.isCancelled {
                cont.yield(produceValue())
                try await Task.sleep(for: .seconds(1))
            }
        }
    }

    cont.finish()
}

Key points:

• sending parameter transfers ownership into the function
• consume continuation allows safe sharing across task group children
• AsyncStream.Continuation is Sendable - no need for nonisolated(unsafe)
• Avoid Task {} inside AsyncStream.init - it's unstructured concurrency

Verification Checklist (When You Change Concurrency Code)

• Confirm build settings (default isolation, strict concurrency, upcoming features) before interpreting diagnostics.
• After refactors:

- Run tests, especially concurrency-sensitive ones (see references/testing.md). - If performance-related, verify with Instruments (see references/performance.md). - If lifetime-related, verify that the code is structured correctly, using with-style APIs.

Glossary

See references/glossary.md for quick definitions of core concurrency terms used across this skill.